1 | //===- SeedCollector.cpp -------------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #include "llvm/Transforms/Vectorize/SandboxVectorizer/SeedCollector.h" |
10 | #include "llvm/Analysis/LoopAccessAnalysis.h" |
11 | #include "llvm/Analysis/ValueTracking.h" |
12 | #include "llvm/IR/Type.h" |
13 | #include "llvm/SandboxIR/Instruction.h" |
14 | #include "llvm/SandboxIR/Utils.h" |
15 | #include "llvm/Support/Compiler.h" |
16 | #include "llvm/Support/Debug.h" |
17 | |
18 | using namespace llvm; |
19 | namespace llvm::sandboxir { |
20 | |
21 | static cl::opt<unsigned> SeedBundleSizeLimit( |
22 | "sbvec-seed-bundle-size-limit" , cl::init(Val: 32), cl::Hidden, |
23 | cl::desc("Limit the size of the seed bundle to cap compilation time." )); |
24 | |
25 | static cl::opt<unsigned> SeedGroupsLimit( |
26 | "sbvec-seed-groups-limit" , cl::init(Val: 256), cl::Hidden, |
27 | cl::desc("Limit the number of collected seeds groups in a BB to " |
28 | "cap compilation time." )); |
29 | |
30 | ArrayRef<Instruction *> SeedBundle::getSlice(unsigned StartIdx, |
31 | unsigned MaxVecRegBits, |
32 | bool ForcePowerOf2) { |
33 | // Use uint32_t here for compatibility with IsPowerOf2_32 |
34 | |
35 | // BitCount tracks the size of the working slice. From that we can tell |
36 | // when the working slice's size is a power-of-two and when it exceeds |
37 | // the legal size in MaxVecBits. |
38 | uint32_t BitCount = 0; |
39 | uint32_t NumElements = 0; |
40 | // Tracks the most recent slice where NumElements gave a power-of-2 BitCount |
41 | uint32_t NumElementsPowerOfTwo = 0; |
42 | uint32_t BitCountPowerOfTwo = 0; |
43 | // Can't start a slice with a used instruction. |
44 | assert(!isUsed(StartIdx) && "Expected unused at StartIdx" ); |
45 | for (Instruction *S : drop_begin(RangeOrContainer&: Seeds, N: StartIdx)) { |
46 | // Stop if this instruction is used. This needs to be done before |
47 | // getNumBits() because a "used" instruction may have been erased. |
48 | if (isUsed(Element: StartIdx + NumElements)) |
49 | break; |
50 | uint32_t InstBits = Utils::getNumBits(I: S); |
51 | // Stop if adding it puts the slice over the limit. |
52 | if (BitCount + InstBits > MaxVecRegBits) |
53 | break; |
54 | NumElements++; |
55 | BitCount += InstBits; |
56 | if (ForcePowerOf2 && isPowerOf2_32(Value: BitCount)) { |
57 | NumElementsPowerOfTwo = NumElements; |
58 | BitCountPowerOfTwo = BitCount; |
59 | } |
60 | } |
61 | if (ForcePowerOf2) { |
62 | NumElements = NumElementsPowerOfTwo; |
63 | BitCount = BitCountPowerOfTwo; |
64 | } |
65 | |
66 | // Return any non-empty slice |
67 | if (NumElements > 1) { |
68 | assert((!ForcePowerOf2 || isPowerOf2_32(BitCount)) && |
69 | "Must be a power of two" ); |
70 | return ArrayRef<Instruction *>(&Seeds[StartIdx], NumElements); |
71 | } |
72 | return {}; |
73 | } |
74 | |
75 | template <typename LoadOrStoreT> |
76 | SeedContainer::KeyT SeedContainer::getKey(LoadOrStoreT *LSI) const { |
77 | assert((isa<LoadInst>(LSI) || isa<StoreInst>(LSI)) && |
78 | "Expected Load or Store!" ); |
79 | Value *Ptr = Utils::getMemInstructionBase(LSI); |
80 | Instruction::Opcode Op = LSI->getOpcode(); |
81 | Type *Ty = Utils::getExpectedType(V: LSI); |
82 | if (auto *VTy = dyn_cast<VectorType>(Val: Ty)) |
83 | Ty = VTy->getElementType(); |
84 | return {Ptr, Ty, Op}; |
85 | } |
86 | |
87 | // Explicit instantiations |
88 | template SeedContainer::KeyT |
89 | SeedContainer::getKey<LoadInst>(LoadInst *LSI) const; |
90 | template SeedContainer::KeyT |
91 | SeedContainer::getKey<StoreInst>(StoreInst *LSI) const; |
92 | |
93 | bool SeedContainer::erase(Instruction *I) { |
94 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) && "Expected Load or Store!" ); |
95 | auto It = SeedLookupMap.find(Val: I); |
96 | if (It == SeedLookupMap.end()) |
97 | return false; |
98 | SeedBundle *Bndl = It->second; |
99 | Bndl->setUsed(I); |
100 | return true; |
101 | } |
102 | |
103 | template <typename LoadOrStoreT> void SeedContainer::insert(LoadOrStoreT *LSI) { |
104 | // Find the bundle containing seeds for this symbol and type-of-access. |
105 | auto &BundleVec = Bundles[getKey(LSI)]; |
106 | // Fill this vector of bundles front to back so that only the last bundle in |
107 | // the vector may have available space. This avoids iteration to find one with |
108 | // space. |
109 | if (BundleVec.empty() || BundleVec.back()->size() == SeedBundleSizeLimit) |
110 | BundleVec.emplace_back(std::make_unique<MemSeedBundle<LoadOrStoreT>>(LSI)); |
111 | else |
112 | BundleVec.back()->insert(LSI, SE); |
113 | |
114 | SeedLookupMap[LSI] = BundleVec.back().get(); |
115 | } |
116 | |
117 | // Explicit instantiations |
118 | template LLVM_EXPORT_TEMPLATE void SeedContainer::insert<LoadInst>(LoadInst *); |
119 | template LLVM_EXPORT_TEMPLATE void |
120 | SeedContainer::insert<StoreInst>(StoreInst *); |
121 | |
122 | #ifndef NDEBUG |
123 | void SeedContainer::print(raw_ostream &OS) const { |
124 | for (const auto &Pair : Bundles) { |
125 | auto [I, Ty, Opc] = Pair.first; |
126 | const auto &SeedsVec = Pair.second; |
127 | std::string RefType = dyn_cast<LoadInst>(I) ? "Load" |
128 | : dyn_cast<StoreInst>(I) ? "Store" |
129 | : "Other" ; |
130 | OS << "[Inst=" << *I << " Ty=" << Ty << " " << RefType << "]\n" ; |
131 | for (const auto &SeedPtr : SeedsVec) { |
132 | SeedPtr->dump(OS); |
133 | OS << "\n" ; |
134 | } |
135 | } |
136 | OS << "\n" ; |
137 | } |
138 | |
139 | LLVM_DUMP_METHOD void SeedContainer::dump() const { print(dbgs()); } |
140 | #endif // NDEBUG |
141 | |
142 | template <typename LoadOrStoreT> static bool isValidMemSeed(LoadOrStoreT *LSI) { |
143 | if (!LSI->isSimple()) |
144 | return false; |
145 | auto *Ty = Utils::getExpectedType(V: LSI); |
146 | // Omit types that are architecturally unvectorizable |
147 | if (Ty->isX86_FP80Ty() || Ty->isPPC_FP128Ty()) |
148 | return false; |
149 | // Omit vector types without compile-time-known lane counts |
150 | if (isa<ScalableVectorType>(Ty)) |
151 | return false; |
152 | if (auto *VTy = dyn_cast<FixedVectorType>(Ty)) |
153 | return VectorType::isValidElementType(ElemTy: VTy->getElementType()); |
154 | return VectorType::isValidElementType(ElemTy: Ty); |
155 | } |
156 | |
157 | template bool isValidMemSeed<LoadInst>(LoadInst *LSI); |
158 | template bool isValidMemSeed<StoreInst>(StoreInst *LSI); |
159 | |
160 | SeedCollector::SeedCollector(BasicBlock *BB, ScalarEvolution &SE, |
161 | bool CollectStores, bool CollectLoads) |
162 | : StoreSeeds(SE), LoadSeeds(SE), Ctx(BB->getContext()) { |
163 | |
164 | if (!CollectStores && !CollectLoads) |
165 | return; |
166 | |
167 | EraseCallbackID = Ctx.registerEraseInstrCallback(CB: [this](Instruction *I) { |
168 | if (auto SI = dyn_cast<StoreInst>(Val: I)) |
169 | StoreSeeds.erase(I: SI); |
170 | else if (auto LI = dyn_cast<LoadInst>(Val: I)) |
171 | LoadSeeds.erase(I: LI); |
172 | }); |
173 | |
174 | // Actually collect the seeds. |
175 | for (auto &I : *BB) { |
176 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: &I)) |
177 | if (CollectStores && isValidMemSeed(LSI: SI)) |
178 | StoreSeeds.insert(LSI: SI); |
179 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: &I)) |
180 | if (CollectLoads && isValidMemSeed(LSI: LI)) |
181 | LoadSeeds.insert(LSI: LI); |
182 | // Cap compilation time. |
183 | if (totalNumSeedGroups() > SeedGroupsLimit) |
184 | break; |
185 | } |
186 | } |
187 | |
188 | SeedCollector::~SeedCollector() { |
189 | Ctx.unregisterEraseInstrCallback(ID: EraseCallbackID); |
190 | } |
191 | |
192 | #ifndef NDEBUG |
193 | void SeedCollector::print(raw_ostream &OS) const { |
194 | OS << "=== StoreSeeds ===\n" ; |
195 | StoreSeeds.print(OS); |
196 | OS << "=== LoadSeeds ===\n" ; |
197 | LoadSeeds.print(OS); |
198 | } |
199 | |
200 | void SeedCollector::dump() const { print(dbgs()); } |
201 | #endif |
202 | |
203 | } // namespace llvm::sandboxir |
204 | |